Paper deep dive

Small-world Networks of Agents Brainstorm AI Risks to Support Ideation

Ke Zhou, Edyta Bogucka, Daniele Quercia

Year: 2026Venue: arXiv preprintArea: cs.HCType: PreprintEmbeddings: 88

Intelligence

Status: not_run | Model: - | Prompt: - | Confidence: 0%

Entities (0)

No extracted entities yet.

Relation Signals (0)

No relation signals yet.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:The ideation phase of participatory AI risk assessment often starts with a blank slate or a limited list of predefined risks, making it difficult to surface indirect or systemic harms. To address this limitation, we propose a three-stage ideation support tool. The tool complements participatory AI, rather than replacing it, and helps focus later engagement with affected communities. First, it dynamically discovers stakeholders depending on the given AI use and recursively expanding outward, allowing overlooked or indirect stakeholders to emerge. Second, it simulates these stakeholders with LLMs, connecting them into a network of a given topology, and having them ideate about risks. Third, it prioritizes risks using network centrality measures. In an initial evaluation, we found that betweenness centrality run through agents connected in a small-world network works best as it elevates risks raised by stakeholders who bridge disconnected groups, surfacing novel, systemic harms that traditional methods often miss. On an AI chatbot companion use case, this approach increased the novelty of the identified risks by approximately 1.1 points over single LLM brainstorming, and by 0.5 points over agentic LLM brainstorming, measured on a normalized five-point Likert scale, without reducing the plausibility or severity of the identified risks. To test whether our framework helps a human-led ideation session using the Futures Wheel approach, we divided 11 teams of non-western young chatbot users into two types: control (team) and treatment (team) in a participatory AI risk assessment. The control teams started from a list of risks generated by the 45 AI practitioners in the initial evaluation; the treatment teams started from a list generated by our framework. The treatment teams identified more risks overall, and more systemic, human-computer interaction, and environmental risks.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

84,338 characters extracted from source content.

Expand or collapse full text

Small-world Networks of Agents Brainstorm AI Risks Ke Zhou 1,2 Edyta Bogucka 1 , Daniele Quercia 1,3 1 Nokia Bell Labs, UK 2 University of Nottingham, UK 3 Politecnico di Torino, Italy ke.zhou@nokia-bell-labs.com, edyta.bogucka@nokia-bell-labs.com, quercia@cantab.net Abstract The ideation phase of participatory AI risk assessment often starts with a blank slate or a limited list of predefined risks, making it difficult to surface indirect or systemic harms. To address this limitation, we propose a three-stage tool to sup- port the ideation phase of AI risk assessment. The tool com- plements participatory AI, rather than replacing it, and helps focus later engagement with affected communities. First, it dynamically discovers stakeholders depending on the given AI use and recursively expanding outward, allowing over- looked or indirect stakeholders to emerge. Second, it simu- lates these stakeholders with LLMs (creating in-silico stake- holders), connecting them into a network of a given topol- ogy, and having them ideate about risks. Third, it prioritizes risks using network centrality measures (the centrality of in- silico stakeholders in the network). In an initial evaluation, we found that betweenness centrality run through agents con- nected in a small-world network works best as it elevates risks raised by stakeholders who bridge disconnected groups, sur- facing novel, systemic harms that traditional methods often miss. On an AI chatbot companion use case, this approach increased the novelty of the identified risks by approximately 1.1 points over single LLM brainstorming, and by 0.5 points over agentic LLM brainstorming, measured on a normalized five-point Likert scale, without reducing the plausibility or severity of the identified risks. To test whether our framework helps a human-led ideation session using the Futures-Wheel approach, we divided 11 teams of non-western young chatbot users into two types: control (team) and treatment (team) in a participatory AI risk assessment. The control teams started from a list of risks generated by the 45 AI practitioners in the initial evaluation; the treatment teams started from a list generated by our framework. The treatment teams identified more risks overall, and more systemic, human-computer in- teraction, and environmental risks in particular, categories that practitioners typically find harder to spot. Project — https://social-dynamics.net/ai-risks/small-world 1 Introduction As AI systems are increasingly integrated into critical soci- etal infrastructure, from healthcare diagnostics to compan- ion chatbots for vulnerable populations, the imperative for comprehensive impact and risk assessment has transitioned from a theoretical ideal to a governance necessity. A crit- ical but often under-examined stage in this process is risk ideation: the initial step in which participants attempt to enumerate as many relevant risks as possible before pri- oritization and mitigation. Because all downstream analy- sis depends on this initial pool, the quality and breadth of ideation largely determine the effectiveness of the entire as- sessment. In practice, however, ideation frequently begins from a blank slate or a limited, organizer-defined list of risks. This creates a structural bottleneck. While frameworks like the EU AI Act (European Commission 2026) emphasize the need to catch “systemic risks”, participants tend to anchor on familiar, high-frequency risks (e.g., privacy or security), while overlooking indirect, systemic, or stakeholder-specific harms (Lancaster et al. 2024; Xia et al. 2023; Fr ̈ ohling et al. 2026). This limitation is not merely a matter of scale, but of epistemic diversity: risks that emerge from the interac- tion between stakeholders, or that disproportionately affect marginalized groups, are systematically underrepresented (Birhane et al. 2022; Lancaster et al. 2024). Traditional human workshops offer depth but are unscal- able and constrained by the “bounded rationality” of small groups of participants (Diehl and Stroebe 1987). Conversely, automated tools that use Large Language Models (LLMs) to generate risk lists, such as Farsight (Wang et al. 2024), ExploreGen (Herdel et al. 2024), and AHA! (Buc ̧inca et al. 2023), act as single expert oracles. They tend to converge on “consensus” risks: those that are plausible and severe, but of- ten generic. Consequently, they struggle to explore the long tail of socio-technical failures that only emerge through the friction of conflicting perspectives, i.e., risks that often carry systemic consequences (Fr ̈ ohling et al. 2026) (e.g., develop- ing an unhealthy over-reliance on the chatbot that hinders real-life social relationship development). In short, current methods are good at confirming what we already know (high likelihood), but less effective at discovering what we do not (high novelty). As a result, they fail to adequately support the human ideation process. To address this, we propose an ideation tool designed specifically to overcome the “blank slate” bottleneck in AI risk assessment. The tool seeds brainstorming workshops with a network of LLM agents, surfacing more diverse and systemic risks than human sessions seeded with practitioner- generated risks. We make three main contributions: arXiv:2609.24859v1 [cs.HC] 21 Sep 2026 Stage 1: Identifying Stakeholders DynamicallyStage 2: Building a Stakeholder Network and Generating Risks Creating an initial set of stakeholders Typical oror Stage 3: Prioritizing Risks Using Network Centrality Degree centralityBetweenness centrality Closeness centrality Ranking the top-K risks from the selected by: Lists of N risks Discovering new stakeholders through recursive role-discovery Lists of N risks A B C Plausibility Likelihood SeverityUniqueness Usability & Risk ideation through Futures Wheel 244 risks 204 risks Random Developer Designer Decision-maker End user Regulator Nuanced System architect System owner Enterprise risk manager Academic researcher Ethics expert recommends Developer System owner Ethics expert recommends vs.FixedSnowballingvs. – 1 5 – 1 5 + 6N – – 1 30 – 1 30 + 31N – 1 2 3 4 5 6 7 8 9 10 Stakeholder roles instantiated as an LLM-based agent HierarchicalSmall-World 1 2 6 3 4 5 7 8 9 10 3 4 2 1 5 6 10 Influence or recommendation relationships Lists of N risksLists of N risks FRAMEWORK FOR GENERATING RISKS Brainstorming Human-Generated Risks in Workshops with a Blank Slate EVALUATION OF THE FRAMEWORK FOR GENERATING RISKS Single-Turn LLM Expert Prompting Multi-Turn Plurals without Network Awareness Risks generated by simulated networks: Risks generated by baselines: EVALUATION OF THE FRAMEWORK FOR SUPPORTING IDEATION IN PARTICIPATORY AI RISK ASSESSMENT 7 8 9 Random Evaluating System-Generated Risks Using Quality Metrics Novelty score Overall score combining Hierarchical Small-World M1 Mapping Human-Generated Risks to System-Generated Risks 45 AI practitioners M2 mapped to Evaluating Human-Generated Risks Quantity of risks M1 Categories of risks M2 Brainstorming Human-Generated Risks in Workshops To Support Ideation 40 AI chatbot users assigned to one of two conditions assigned to one of two conditions Reading an risk assessment report with: based on provided risk | ControlPractitioner-generated risks C1 | Treatment Network-generated risks C2 from non-Western countries generated each risk’s: Quality metrics Figure 1: Overview of the proposed framework and its two evaluations. (A) Our three-stage framework for generating risks. It involves identifying diverse stakeholders dynamically (Stage 1); building a stakeholder of in-silico agents network and generating risks by orchestrating the communication through a network with a given topology (Stage 2); and prioritizing risks using three centrality metrics: degree, betweenness, and closeness (Stage 3). (B) Evaluation of the framework’s risk generation quality by comparing system-generated risks with practitioners-generated risks from workshops. (C) Evaluation of the framework as a support tool for participatory AI risk assessment. Participants are assigned to either a control condition (report with practitioners-generated risks) or a treatment condition (report with network-generated risks), and use the provided reports as starting points for Futures Wheel ideation. (1) A Three-Stage Framework for Networked Stakeholder Simulation (Section 3): We propose a framework that dy- namically discovers stakeholders via recursive snowballing, instantiates them as in-silico LLM agents, and connects them into a network topology through which they commu- nicate to brainstorm risks. Risks are then prioritized using network centrality measures applied to the resulting stake- holder graph. We show small-world topologies combined with betweenness centrality are uniquely effective at surfac- ing novel, systemic harms that flat ideation methods miss. (2) Evaluation of Risk Generation Quality (Section 4): We evaluate our framework against single-LLM and multi-agent baselines using a dataset of 244 risks generated by 45 AI practitioners in blank-slate brainstorming workshops. Our approach improves novelty by approximately 1.1 points over single-LLM brainstorming, and 0.5 points over agentic LLM baselines, without reducing plausibility or severity. (3) Evaluation of Support for Human-Led Participatory AI Risk Assessment (Section 5): We conduct a controlled Fu- tures Wheel study (Glenn 2009) with 40 non-Western young chatbot users. Treatment teams seeded with our frame- work’s risks identified substantially more risks overall, and significantly more systemic, human-computer interaction, and socioeconomic harms than control teams seeded with practitioner-generated risks. 2 Related Work 2.1 AI Risk Identification and Assessment AI risk identification and AI risk assessment are distinct but related processes: AI risk identification uncovers potential harms, while risk assessment quantifies them for prioritiza- tion and decision-making. AI risk identification has evolved from narrowly defined technical safety evaluations toward broader socio-technical analyses. Early approaches surfaced technical risks at the model level, such as robustness fail- ures and security vulnerabilities, and quantified them using metrics including error rates, perturbation robustness, and adversarial vulnerability (Schmitz et al. 2025; Steimers and Schneider 2022). As AI systems became embedded in social and institutional contexts, research attention shifted toward identifying and quantifying downstream harms. AI risk taxonomies have played a central role in this shift. Influential work by Weidinger et al. (2022), alongside sub- sequent syntheses (Uuk et al. 2024; Zhang et al. 2025), pro- vides comprehensive catalogues of risks spanning techni- cal failures, human–AI interaction risks, and long-term sys- temic societal risks. These taxonomies offer shared vocab- ulary and conceptual clarity, but they are largely retrospec- tive: they systematize known risks rather than supporting the anticipation of novel, context-specific risks in emerging ap- plications. Moreover, they typically abstract away from the perspectives of distinct stakeholders. Participatory approaches attempt to address these limita- tions by incorporating diverse human perspectives into risk identification (Delgado et al. 2023; Lee et al. 2019). Meth- ods such as the Judgment Call card game (Ballard, Chappell, and Kennedy 2019), participatory scenario planning (Kies- lich, Helberger, and Diakopoulos 2026), and end users audit- ing (Deng et al. 2025) leverage structured interaction to sur- face values, concerns, and experiential knowledge that may otherwise remain implicit. While effective in eliciting rich insights, these approaches are resource-intensive, difficult to scale, and constrained by practical barriers to participation. In particular, assembling representative groups of end users, regulators, and marginalized stakeholders is often infeasible for organizations conducting AI risk assessments. In addi- tion, current participatory AI literature (Kallina, Bohn ́ e, and Singh 2025; Birhane et al. 2022; Corbett, Denton, and Erete 2023; Delgado et al. 2023) overlooks the “blank slate” bot- tleneck in risk ideation, where forcing participants to brain- storm from scratch exhausts their cognitive bandwidth on obvious issues (Hohma et al. 2023). Current AI risk assessment practices are strongly influ- enced by regulatory frameworks (European Commission 2026), and predominantly focus on evaluating identified risks in terms of likelihood and severity, an approach in- herited from risk assessment traditions in domains such as safety engineering and finance. These dimensions are also prominent evaluation criteria in prior research (Wang et al. 2024). By contrast, novelty and diversity of perspectives have rarely been treated as explicit assessment objectives (Herdel et al. 2024), and there remains a lack of robust mea- sures for how they should be defined and operationalized. 2.2 AI-Augmented Brainstorming for Risk Assessment The use of AI systems to support brainstorming and ideation has gained increasing attention (Wang et al. 2025; Ashk- inaze et al. 2025b; Choi et al. 2024). Empirical studies show that LLMs can substantially increase the quantity of generated ideas (Bouschery, Blazevic, and Piller 2024; Yu- Han and Chun-Ching 2023). However, these gains are fre- quently accompanied by reductions in diversity and origi- nality, with models exhibiting fixation on stereotypical or high-frequency patterns (Wadinambiarachchi et al. 2024; Meincke, Nave, and Terwiesch 2025; AlDahoul, Rahwan, and Zaki 2025; Bai et al. 2025). This tendency is par- ticularly salient in risk assessment contexts. For example, when prompted to identify “AI risks”, LLMs often default to generic concerns such as privacy or bias, overlooking sub- tle, domain-specific, or interactional harms (Lancaster et al. 2024; Xia et al. 2023; Fr ̈ ohling et al. 2026). To scale risk discovery, the industry has increasingly turned to AI red teaming and AI-augmented brainstorm- ing. Red teaming leverages adversarial testing to probe mod- els for vulnerabilities and policy violations (Ge et al. 2024; Ganguli et al. 2022). While automated red teaming is highly effective at identifying technical exploits (e.g., jailbreaks), it fundamentally relies on predefined threat models and often struggles to anticipate nuanced, socio-technical harms that only emerge through complex stakeholder interactions. To broaden the scope of risk identification beyond red team- ing, several tools leverage LLMs to assist AI risk assess- ment directly. Systems such as Farsight (Wang et al. 2024), ExploreGen (Herdel et al. 2024), and AHA! (Buc ̧inca et al. 2023) generate or visualize candidate risks based on system descriptions. While these tools improve efficiency, they typ- ically treat the LLM as a single expert oracle. As a result, they do not explicitly model interaction, disagreement, or perspective-taking among stakeholders, mechanisms that are known to underpin diversity in human group brainstorming (Meincke, Nave, and Terwiesch 2025). 2.3 Stakeholder Analysis and Network Science Recent work has explored the use of LLMs to simulate aspects of human behavior and social interaction, often described as in-silico social science. The Plurals frame- work (Ashkinaze et al. 2025a) demonstrates that LLM agents initialized with distinct roles or personas can par- ticipate in structured deliberative processes, producing out- puts that differ meaningfully from single-agent prompting. Similarly, Park et al.’s Generative Agents (Park et al. 2023) show that agents endowed with memory and social con- text can exhibit emergent social dynamics. Within AI risk identification, however, prior applications of multi-agent LLM systems have largely relied on unstructured debate or flat ideation structures. Fukumura and Ito (2025) explored agent-supported brainstorming, and found that benefits were primarily limited to general or non-specialized topics. These findings suggest that role differentiation alone may be insuf- ficient; the structure of interaction itself plays a critical role in shaping outcomes. Stakeholder theory posits that organizational outcomes depend on the relationships among stakeholders, not solely on individual attributes (Parmar et al. 2010). Network sci- ence provides formal tools for analyzing these relationships. Rowley (Rowley 1997) argues that stakeholder influence is a function of network position such as centrality or brokerage rather than intrinsic importance alone. In adjacent domains, network measures (Borgatti, Jones, and Everett 1998; Watts and Strogatz 1998) have been used to identify influential ac- tors, information bottlenecks, and structural vulnerabilities. However, these methods have rarely been applied to in-silico brainstorming or assessment processes. Existing approaches typically aggregate risks by frequency or expert judgment, implicitly assuming all contributors are equally informative. Research Gap. The initial ideation phase often starts from a “blank slate” or a narrow set of predefined risks. This forces participants to spend time listing well-known concerns be- fore they can address broader, systemic risks. To address this gap, we introduce a tool that helps generate early ideas about AI risks by simulating discussions among in-silico stake- holders. The tool structures these interactions as networks, allowing us to explore how communication patterns shape the risks they identify. This approach surfaces risks from a wider range of perspectives, especially from stakeholders who “bridge” otherwise separate groups. 3 Our Framework Our framework operationalizes AI risk brainstorming through network-mediated ideation, a process where in- silico agents representing diverse stakeholders generate risks within a network topology. Rather than treating risk identi- fication as an isolated ideation task, we conceptualize it as a collaborative activity shaped by who participates, how they are connected, and whose perspectives bridge across roles. Figure 1 provides an overview of the methodology, con- sisting of three stages: (1) Discovering stakeholders dynami- cally (Section 3.1); (2) Building a stakeholder network (Sec- tion 3.2); and (3) Prioritizing risks using network centrality (Section 3.3). Formally, given an AI use case, our goal is to produce a set of top-K risks R = r 1 ,r 2 ,...,r K on a ranked list, where each risk r k is associated with one or more stakeholders. 3.1 Stage 1: Identifying Stakeholders Dynamically Creating An Initial Set of Stakeholders. We begin by iden- tifying an initial set of stakeholders relevant to the AI use case U :S 0 = s 1 ,s 2 ,...,s m , where each s i corresponds to an abstract role rather than a specific individual. We define a typical stakeholder set grounded in respon- sible AI practice and governance literature (San Vito et al. 2025). This set includes five roles: Developers, responsi- ble for building and maintaining the AI system; Designers, shaping user interaction and experience; Decision-makers, responsible for deployment and organizational strategy; End users, who directly interact with or are affected by the sys- tem; and Regulators, responsible for oversight and compli- ance. These roles align with prior work on AI auditing and accountability (San Vito et al. 2025), which emphasizes the inclusion of decision subjects, domain experts, and regula- tory actors in assessment processes. For AI companion chat- bots in particular, these categories capture both technical and socio-emotional dimensions of risk. We also define a nuanced stakeholder set based on NIST standard framework for AI risk assessments (AI 2023), in- cluding AI developers, machine learning engineers, data scientists, data collectors and curators, system architects, model trainers, and technical auditors; product and project managers, system owners, organizational leadership, enter- prise risk managers, legal counsel, compliance officers, pro- curement teams, and IT deployment and operations staff; clinicians, judges, HR professionals, and analysts; external regulators, supervisory authorities, policymakers, and legis- lators; and academic researchers, ethics experts, journalists and environmental stakeholders. Making The Set of Stakeholders Comprehensive Via Snowballing. To capture stakeholders omitted by the initial abstraction, we perform iterative stakeholder expansion by employing an LLM-driven snowballing technique for rec- ommending more relevant stakeholders recursively. At iter- ation t, each stakeholder s i ∈S t proposes a set of additional stakeholders ∆S (t) i = s i1 ,s i2 ,..., where we query the LLM : “Given the AI use case [Description] and your stake- holder role as [Stakeholder], list what other stakeholders (no personal names) are significantly impacted by or have influence over this system?” The candidate set for iteration t + 1 is:S t+1 =S t S i ∆S (t) i . To prevent unbounded growth, we deduplicate roles us- ing semantic similarity. Let φ(s) denote an embedding of a stakeholder description. A newly proposed stakeholder s ′ is merged with an existing stakeholder s if: cos(φ(s ′ ),φ(s))≥ τ , Expansion terminates when either there are no novel stakeholders are added (S t+1 = S t ), or a maximum num- ber of iterations T max is reached. 3.2 Stage 2: Building a Stakeholder Network and Generating Risks Connecting Stakeholders In A Network Within A Given Topology. We define a stakeholder interaction network as a directed weighted graph: G = (V,E,W ), where: V = S T is the final stakeholder set, E ⊆ V × V represents recom- mendation of stakeholders indicting influences, W : E → R + assigns edge weights. An edge (s i ,s j ) exists, if stake- holder s j is recommended by stakeholder s i , and therefore deemed to maintain influence over stakeholders j during risk ideation. Using this stakeholder set, we generate multiple in- stantiations of G with different network topologies: Hierarchical Network: It is constructed by assigning stakeholders to levels ℓ(s) ∈N based on organizational au- thority. Edges are constrained such that: ℓ(s i ) < ℓ(s j ) ⇒ (s i → s j ). This reflects top-down flow typical of formal governance structures. Small-World Network: It is designed to balance dense local interaction among closely related stakeholder roles with a limited number of long-range connections across the network. We construct a small-world–inspired directed acyclic graph by preserving strong high-weight edges within stakeholder clusters identified through the recommendation, while enforcing acyclicity. Formally, an edge is retained be- tween stakeholders s i ,s j ∈ V , if one recommends the other, and the edge does not form circles. This structure supports locally coherent ideation, while enabling risks to propagate across stakeholder communities through bridging roles. Random Network: In the random baseline, edges are sam- pled uniformly at random from V × V , subject to acyclicity and fixed edge count|E|. Pruning Network Graph and Removing Cycles. Given the directed interaction structures, we transform G into a di- rected weighted graph G ′ by performing backbone extrac- tion, and graph cycle removal procedures. Backbone Extraction. The initial stakeholder interaction graph G = (V,E,W ) may contain a dense set of edges reflecting weak, redundant, or noisy relationships. We apply a disparity filter (Serrano, Bogun ́ a, and Vespignani 2009) to remove statistically insignificant edges, retaining only edges whose normalized weight exceeds a significance thresholdδ. We set δ = 0.05 following prior work on backbone extrac- tion in weighted networks (Serrano, Bogun ́ a, and Vespignani 2009). Applying the disparity filter yields a pruned graph: G b = (V,E b ,W b ), where E b ⊆ E contains only statisti- cally significant edges. This backbone graph retains the ma- jority of total edge weight, while substantially improving in- terpretability and computational efficiency. Graph Cycle Removal. The network-mediated ideation structures are required to be directed acyclic graphs (DAGs) in order to support multi-round, feed-forward information flow. However, the backbone graph G b may still contain cy- cles, reflecting reciprocal or mutual influence among stake- holders. To enforce acyclicity, we therefore transform G b into a DAG G ′ using a greedy maximum-weight acyclic subgraph construction. Given a directed weighted graph G b = (V,E b ,W b ), we seek a directed acyclic subgraph G ′ = (V,E ′ ,W ′ ) such that: E ′ ⊆ E b , and the total retained weight P (i,j)∈E ′ w ij is maximized. This problem is NP- hard in general. We therefore adopt an efficient greedy ap- proximation: sorting remaining edges by descending weight, and greedily adding them to G ′ unless doing so would create a cycle. This algorithm prioritizes preserving high-weight influence relationships while eliminating weaker reciprocal ties that would violate acyclicity. As a result, G ′ approxi- mates the most influential pathways within the stakeholder system. To assess robustness, we verified that alternative cycle-breaking strategies (e.g., random edge removal) pro- duce qualitatively similar results in downstream risk diver- sity and ranking, though with slightly higher variance. Generating Risks By Orchestrating The Communica- tion Among Stakeholders. Each stakeholder s i ∈ V is instantiated as an in-silico agent a i , parameterized by: a i = ⟨s i ,U,I (t) i ⟩, where I (t) i denotes information avail- able at round t for the use case U . Stakeholder agents generate risks via conditional language modeling: r (t) ik ∼ p θ (r | s i ,U,I (t) i ), where θ denotes the underlying LLM parameters. At each round t, stakeholder agent a i re- ceives inputs from its immediate predecessors: I (t) i = S (s j →s i )∈E ′ R (t−1) j , whereR (t−1) j is the set of risks gener- ated by stakeholder agent a j in the previous round. This in- duces a structured dependency between stakeholder agents’ outputs, operationalizing network-mediated ideation. 3.3 Stage 3: Prioritizing Risks Using Network Centrality Let G ′ = (V,E ′ ) be the final stakeholder graph on which we compute three network measures. A core hypothesis of the work is that who identifies a risk matters as much as what the risk is. We derive three primary rankings for prioritizing risks based on these network centrality measures: Degree Centrality Ranking. Risks are ranked by the sum of the Degree Centrality of the stakeholders who identified them: Score D (R) = P s∈Contributors(R) C deg (s). This fa- vors risks identified by “popular” or most obvious stakehold- ers (e.g., Developers, Product Managers). Betweenness Centrality Ranking. Risks are ranked by the sum of the Betweenness Centrality of the stakeholders: Score B (R) = P s∈Contributors(R) C bet (s). Betweenness centrality (C bet ) measures how often a node acts as a bridge along the shortest path between two other nodes. In our con- text, high-betweenness stakeholders are hypothesized to see risks that arise from stakeholders that bridge disconnected groups (e.g., “Ethics & Compliance Officer” sitting between “Regulatory Body” and “Developer”). ClosenessCentralityRanking.Risksareranked by the sum of the Closeness Centrality of the stakeholders who identified them: Score C (R)= P s∈Contributors(R) C clo (s). Closeness centrality (C clo ) captures how close a stakeholder is, on average, to all other stakeholders in the network, based on shortest path distances. This ranking strategy prioritizes risks raised by stakeholders who are well-positioned to rapidly inte- grate information across the network, favoring broadly informed perspectives rather than highly popular (degree) or structurally bridging (betweenness) roles. In summary, these three ranking strategies reflect distinct theories of influence in collective risk assessment: degree centrality emphasizes visibility and consensus, betweenness centrality highlights cross-role mediation and systemic in- sight, and closeness centrality prioritizes globally informed perspectives. Next, we empirically evaluate our proposed framework on an AI chatbot companion use case. We selected this specific use case because it represents a high-stakes, socio-emotional application characterized by wide demographic use and timely regulatory attention. We evaluate our framework in two ways. First, we evaluate our framework with blank-slate of risks by AI practitioners (Section 4). Second, we directly test our framework’s effectiveness in supporting human-led ideation in Participatory Risk Assessments through a con- trolled study using the Futures Wheel technique (Glenn 2009) (Section 5), in which non-Western young chatbot users who started from risks generated by our framework are compared to those starting from practitioners-generated risks in our first evaluation (i.e., by AI practitioners). 4 Evaluation of Our Framework for Generating Risks To validate the efficacy of our framework, we evaluate whether networked simulation improves risk generation compared to flat baselines without network structures, using a dataset of practitioners-generated risks as a benchmark. 4.1 Establishing Brainstorming Workshops for Generating Practitioners-based Risks To evaluate our framework, we drew on a dataset from 8 human-based brainstorming workshops focused on identify- ing risks associated with AI chatbot companions (Govers, ˇ S ́ cepanovi ́ c, and Quercia 2026). Rather than representing an idealized participatory engagement, these workshops were designed to mimic the typical, early-stage internal risk work- shops commonly conducted within public institutions and private companies. We recruited 45 AI practitioners across 3 locations via mailing lists, online ads, and EventBrite. All participants reported active personal use of AI chatbot com- panions, grounding their risk identification in everyday AI chatbot interaction. Those workshops utilized blank slate brainstorming (i.e., with no risks provided), each lasting 1 hour and 45 minutes with 4 to 7 participants. The study received institutional ethics approval and participants were compensated above minimum wage. Participants (detailed demographics shown in Appendix) were relatively young (majority aged 25–34), highly educated and all held univer- sity degrees (51% Master’s, 20% PhD). Many participants came from institutions and companies (37% were from the company, 43% researchers, and 20% students). To capture a rich socio-technical perspective, the team specifically in- cluded: (1) AI professionals, comprising software develop- ers, NLP experts, AI monitoring/auditing experts, and de- signers, who contributed insights into technical and opera- tional failure modes; (2) Applied AI researchers, who devel- oped AI algorithms in a company and evaluated their limi- tations; (3) Researchers in Responsible AI and psychology, who specialized in the socio-emotional and downstream im- pacts of companions; and (4) Civil society representatives, who advocated for protection of vulnerable groups. Self- reported AI expertise was high, with 65% of participants re- porting knowledgeable or expert levels. All workshops were balanced to include company workers, researchers, and stu- dents. Across all sessions, participants generated 244 dis- tinct risks, which form the basis of our evaluation. 4.2 Defining the Multi-Dimensional Rubric We adopted a multi-dimensional rubric (Table 1) to evalu- ate the quality of practitioners-generated risks for chatbot companion within those brainstorming workshops. Synthe- sizing criteria from the NIST AI Risk Management Frame- work and other prior AI risk assessment frameworks (AI 2023; Diakopoulos and Johnson 2021; Kieslich, Helberger, and Diakopoulos 2026; Dean et al. 2006; Lewis and Sauro 2009; Brooke et al. 1996), the rubric assesses plausibil- ity, probability, severity, uniqueness, novelty, and usability. To explicitly distinguish between evaluation constructs, our rubric separates Plausibility (theoretical feasibility), Like- lihood (realistic probability of occurrence in deployment), and Uniqueness (specificity across domains), a correlation matrix provided in Appendix demonstrates that these dimen- sions capture distinct evaluative signals rather than redun- dant measures. Five external AI-domain experts recruited via Prolific rated the risks using Likert scales, following at- tention and comprehension screening. They are AI ethicists Dimension Definition Plausibility Assesses whether the described risk could realisti- cally arise in the given AI use context (1–5 Likert (Diakopoulos and Johnson 2021)). Likelihood Estimates the likelihood that a risk will occur (1– 7 Likert, based on NIST AI risk assessment mea- sures (AI 2023) and foresight (Fr ̈ ohling et al. 2026)). Uniqueness Evaluates whether the risk is specific to the particu- lar AI use case, rather than broadly applicable across AI systems or domains (1–3 Likert (Kieslich, Hel- berger, and Diakopoulos 2026)). NoveltyRates the degree of creativity and non-obviousness of the risk, ranging from mundane or checklist-like to imaginative and original (1–3 Likert (Dean et al. 2006; Fr ̈ ohling et al. 2026), ). UsabilityMeasures how clear, understandable, and practically engaging the risk (1–5 Likert (Lewis and Sauro 2009; Brooke et al. 1996)). SeverityAssesses the seriousness of potential harm and its downstream social, economic, or institutional conse- quences (1–5 Likert (AI 2023; Fr ̈ ohling et al. 2026)). Table 1: Evaluation rubric used for annotating AI risks. Each dimension corresponds to explicit annotation questions posed to evaluators, and is grounded in prior literature. who frequently develop assessments for AI. For consistency and ease of comparisons, all the ratings were normalized to five-point Likert scale. These expert annotations provide a quantitative basis for comparing risk quality across brain- storming methods and with system-generated outputs. 4.3 Establishing Single-/Multi-Agent Baselines We compare our approach against two baselines commonly used in AI risk assessment: Single-Turn LLM Expert Prompting: A single LLM is prompted to generate and rank risks from an expert perspec- tive, conditioned on a description of the AI use case sys- tem and a list of stakeholders. This single LLM approach reflects current practice in automated risk generation tools, such as Farsight (Wang et al. 2024), ExploreGen (Herdel et al. 2024), and AHA! (Buc ̧inca et al. 2023). Multi-Turn Plurals without Network Awareness: We use standard Plurals framework (Ashkinaze et al. 2025a) to sim- ulate multiple in-silico stakeholder agents interacting in a flat ensemble structure, without explicit network structure (topology) or network-based ranking . This baseline isolates the contribution of network structure and measures, and rep- resents current agentic LLM brainstorming method. To evaluate those two baselines and our framework, we mapped all system-generated risks to practitioners- generated risks using a two-stage process combining embedding-based similarity and manual validation. Each of the 244 practitioners-curated risks was embedded and matched to the five most similar system outputs, producing over 1,000 candidate pairs. We manually validated semantic matches, allowing one-to-many mappings. In total, 108 hu- man risks (44%) matched at least one system-generated risk. These matched risks were used for our evaluation. NETWORK TOPOLOGYCENTRALITY MEASURES A B BASELINES Metric (normalized five-point Likert scale) 0 1 2 3 4 5 Novelty score Combined score Single-turn LLM Multi-turn Plurals Hierarchical Network Random Network Small-World Network Degree centrality Closeness centrality Betweenness centrality 3.29 2.32 3.72 2.98 3.73 2.95 3.90 3.05 4.05 3.18 4.05 3.18 3.02 3.96 3.99 3.48 0 1 2 3 4 5 VS. Figure 2: Our analysis utilizes two metrics: “Combined Score” aggregates Plausibility, Likelihood, Uniqueness, Usability and Severity while “Novelty” reflects risk diversity. All measures were normalized to the five-point Likert scale, and the top per- forming systems are bolded. A: Comparison of Network Topologies (using Snowballing Typical Roles and Degree Centrality): Small-World networks outperform both structured hierarchies and unstructured random networks. They significantly outper- form the Baselines, demonstrating the value of structured interaction. B: Impact of Centrality Measures on Risk Quality (Small- World Network + Snowballing Typical Roles): Degree Centrality yields the highest quality (Combined Score) by capturing consensus; Betweenness Centrality is most effective at surfacing Novel risks by prioritizing risks from bridging stakeholders. 4.4 Results: Network Prioritization Drives Novelty We evaluate our framework against static and unstructured baselines using two metrics: a Combined Score (an aggre- gate of plausibility, likelihood, uniqueness, usability, and severity), and a Novelty score (specifically isolating risk di- versity and non-obviousness). All scores are normalized to a five-point Likert scale. We explicitly report Novelty score to capture non-obvious risks, often disproportionately affect- ing minority or marginalized populations, that are systemat- ically obscured by consensus-driven aggregates. By moving beyond flat lists of stakeholders, we show that network topology, centrality measures, and discovery strategies fundamentally shape the quality of generated risks (evaluation results on all system variants are provided in Ap- pendix with a detailed discussion of the discovery strate- gies). We highlight our key findings below. Small-World Topologies Optimize Ideation Efficiency. Network science literature (Watts and Strogatz 1998) posits that “small-world” networks, characterized by high local clustering and short average path lengths, are uniquely ef- ficient at information diffusion. We find that this structural advantage holds for simulated AI risk ideation, significantly outperforming unstructured interactions. Figure 2A com- pares three network topologies using the same strategy for creating the set of stakeholders: Small-World network achieves the best balance (Combined 4.05, Novelty 3.18). The high local clusters allows distinct stakeholder groups (e.g., a “Security Cluster” vs. a “Patient Advocacy Cluster”) to develop coherent, deep arguments lo- cally before propagating them. Random network performs competitively on the Combined Score but suffer on Novelty (3.05). We found stakeholders lack the local context to refine specific critiques, leading to more generic outputs. Hierarchical network performs poorly on Novelty (2.95). Designed to mimic corporate reporting lines, these struc- tures appear to introduce information filtering. “Middle manager” agents may inadvertently smooth over radical sig- nals from leaf nodes (end-users) before they reach the cen- tral nodes, effectively suppressing diverse risk identification. The structural impact is evident when comparing Small- World results to the Single-turn LLM (Combined 3.29) and Multi-turn Plurals (Combined 3.72). The Single-turn base- line lacks the iterative refinement of ideas, while the Multi- turn baseline lacks the community structure to support spe- cialized discourse. The Small-World topology provides a +0.76 improvement in Combined Score over the standard LLM baseline, demonstrating that how agents are connected is as critical as the agents themselves. Novelty is Driven by Bridging Stakeholders (Using Be- tweenness Centrality). A core challenge in AI auditing is moving beyond “known knowns” (high-consensus risks) to uncover “unknown unknowns”. Standard methods often converge on high-likelihood but generic risks (DeVos et al. 2022; Lancaster et al. 2024). We find that the choice of net- work centrality measure used to prioritize risks alters the na- ture of the output, revealing a trade-off between consensus and discovery (Figure 2B) Ranking risks by Degree Central- ity (popularity) yields the highest Combined Scores (4.05). However, qualitative analysis suggests these risks often rep- resent consensus views, broadly recognized issues like data privacy leaks or standard bias. While these are valid, they are often already known to development teams. In contrast, prioritizing risks via Betweenness Centrality results in the highest Novelty scores (3.48). Betweenness centrality iden- tifies stakeholders who act as bridges between otherwise dis- connected sub-communities. Unlike high-degree nodes that reinforce dominant narratives, bridging nodes are uniquely positioned to translate context across epistemic boundaries. To illustrate this mechanism, we analyzed the risks gen- erated by the top 10 stakeholders ranked by Betweenness Centrality (Figure 3). Some of those top 10 stakeholders also overlap with those in top 10 of Degree Centrality mea- sures such as IT Manager and Product Manger. The data re- veals a stark divergence in Novelty depending on the type of domain the stakeholder bridges. Stakeholders who also has high Degree Centrality tended to identify valid but low- Ethics Officer Family CaregiverClinical Staff Caregiver Healthcare Compliance Officer Pharmacist Sibling Family member Psychologist Crisis Counselor CS Manager System Administrator IT Support Specialist Technologist IT Support CS Manager Developer Training Specialist TS Specialist CS Agent CS Lead CS Specialist UX/UI Designer Training Coordinator AI Interaction Designer Marketing Manager Content Creator Content Strategist Data Analyst UX Researcher UX Designer QA Tester User Experience Designer Accessibility Lead QA Tester QA Team QA Analyst Customer Experience Manager IT Specialist Tech Expert Tech Lead Software Engineer Software Developer TS Lead Training Lead Project Manager CR Representative Cybersec Analyst Nurse IT Staff QA Specialist Healthcare Provider End User QA Analyst Helpdesk Agent Consumer Advocate CX Analyst Clinical Engineer Health IT Expert Healthcare Administrator Regulatory Authority Data Privacy Officer Regulatory Affairs Specialist Project Sponsor Marketing Officer User IT Security Specialist DP Authority Officer Legal Counsel Privacy Officer Regulatory Body Ethical Review Board Member Ethics and Compliance Officer Legal Advisor AI Ethicist QA Officer Compliance Specialist CP Advocate Compliance Officer CP Officer IT Security Officer DP Officer Policy Maker Clinical Worker Social Worker Health Professional Behavioral Expert Ethicist Health Advocate Counselor Teacher Clinical Psychologist Study Advisor Psychiatrist Healthcare Professional Mental Health Expert Digital Wellbeing Researcher Technology Ethicist Ethics Committee Member Technology Ethics Advisor Peer Student IT Security Manager Behavioral Clinician Data Scientist Insurance Agent QA Quality Assurance TS Technical Support DP Data Protection CP Consumer Protection CR Consumer Rights Product Manager Product Manager Product Manager Shortcuts: A STAKE- HOLDER RISK STATEMENT NOVELTY SCORE Policy Maker User Ex- perience Designer Failure to identify crisis situations or escalate them could endanger user safety during emergencies Mental Health Researcher Failure to replicate evidence- based therapeutic techniques could worsen mental health Chatbot overreliance could hinder relationship development Parent Misinterpreting medication instructions provided by chatbot could lead to non-adherence Clinical Pharmacist Providing inaccurate health responses could cause confusion and adverse patient outcomes Nurse Poor 24/7 scalability could cause crashes and disrupt support IT Manager Failure to recognize emotional cues may make interactions feel impersonal during distress End User Infrastructure incompatibility could delay deployment and reduce operational effectiveness Product Manager Failure to incorporate clinical insights may result in a product that ignores emotional needs Failure to adhere to transparency and bias standards triggers regulatory scrutiny Regulatory Body 4 4 5 4 4 4 1 2 3 3 TOP 10 STAKEHOLDERS BY BETWEENNESS CENTRALITY MEASURE BSTAKEHOLDER NETWORK CS Customer Support Parent Clinical Pharmacist Mental Health Researcher CR Advocate TS IT Manager Patient Educator Student Figure 3: A: Risks and their novelty scores generated by the top 10 stakeholders ranked by Betweenness Centrality. B: The stakeholder network constructed through snowballing, within which in-silico LLM agents are connected and communicate to brainstorm risks. Many stakeholders who bridge diverse domains with top Betweenness Centrality such as Parent, Mental Health Researchers, Policy Makers and Clinical Pharmacist are the primary identifiers of highly novel risks (scores 4–5) such as psychological dependency and user safety during crisis. novelty risks. For example, the Product Manager identi- fied the risk of “Incompatibility with existing infrastruc- ture” (Novelty=1), and the IT Manager identified scalabil- ity risks regarding “24/7 volume causing crashes” (Nov- elty=2). Their outputs reflect standard operational concerns common in software deployment. In contrast, stakeholders who are mainly bridging sub-communities with extremely high Betweenness Centrality but relatively low Degree Cen- trality provided the highest novelty risks: The Parent (Novelty 5), acting as a bridge between the do- mestic sphere and the educational/technical system, identi- fied the risk of “Developing an unhealthy over-reliance on the chatbot”, hindering real-life social development. This is a profound socio-psychological harm that purely technical agents missed, although this harm is documented in prior lit- erature (e.g., loneliness-through-technology (Wilson 2018)). The Mental Health Researcher and Clinical Pharmacist (both Novelty 4) bridged clinical standards and AI interac- tion, identifying risks such as “Inability to effectively repli- cate evidence-based therapeutic techniques” and “Misinter- preting medication instructions”, respectively. The Policy Maker (Novelty 4), bridging public safety and user interaction, highlighted the systemic failure to “identify crisis situations or escalate them”, a critical safety gap. This analysis confirms that novelty is not randomly dis- tributed. It is structurally concentrated in stakeholders who navigate the boundaries between distinct communities. By prioritizing these bridging nodes via Betweenness Central- ity, the system identifies high-impact, non-obvious risks that are invisible to the commonly identified risks. 5 Evaluation of Our Framework Supporting Participatory AI Risk Assessment In this section, we focus on testing whether our framework actually helps a human-led ideation session. The goal of this evaluation was to test whether seeding a participatory ses- sion with network-generated risks leads participants to iden- tify more risks, particularly systemic and downstream risks, than seeding the session with practitioner-generated risks. 5.1 Participatory AI Risk Assessment with Futures Wheel We conducted a user study using the Futures Wheel method (Glenn 2009). It is a structured brainstorming technique for identifying first-, second-, and third-order consequences of a central topic, enabling participants to explore cascad- ing and interconnected effects. Starting from an AI chatbot companion, participants iteratively answer the question “If this occurs, what happens next?”, generating primary conse- quences in the first layer, followed by secondary and tertiary consequences in the second and third layers. This process encourages participants to move beyond immediate risks to- ward more systemic and long-term impacts. We selected the Futures Wheel as our method of partic- ipatory AI risk assessment for three reasons. First, it offers a simple, structured format for team ideation that enables participants to generate a large number of risks without re- quiring domain expertise (Glenn 2009). Second, it pushes participants to think beyond obvious risks for AI chatbots by exploring their second- and third-order consequences (Bengston 2016). Third, it has been successfully used in large-scale participatory AI risk assessment studies (with over 250 participants across six Global Majority and Minor- ity countries (Hohendanner et al. 2025)), making it appro- priate for our intended participants. Participant Demographics and Recruitment. We are par- ticularly interested in testing how our framework supports risk ideation among a crucial but under-represented demo- graphic in AI risk literature: young, frequent chatbot users from predominantly non-Western (non-US, non-European) backgrounds (Png 2022; Birhane et al. 2022). Young users in the Global South and non-Western contexts are major consumers of AI companions (Scherr et al. 2025), yet their perspectives are rarely centered in participatory risk assess- ments (Png 2022; Birhane et al. 2022). We recruited 11 teams of participants, comprising a total of 40 individuals (with three to four people per team). Participants generally reported high AI chatbot usage: with 20 participants (50%) using AI chatbots multiple times per day, 6 (15%) once daily, 11 (27.5%) 2–6 times per week, and 3 (7.5%) once per week. Additionally, 13 participants (32.5%) reported having expe- rienced technology-related harms, citing issues such as mis- information, privacy loss, social media addiction, surveil- lance, and over-reliance on technology. Ethnically, the co- hort was diverse: 15 participants (37.5%) identified as Mid- dle Eastern or North African, 10 (25%) as Asian, 8 (20%) as White, 4 (10%) as Black, 1 (2.5%) as Mixed, 1 (2.5%) as Central Asian, and 1 (2.5%) preferred not to say. Experimental Conditions: Practitioners-Seeded (Con- trol) vs. Framework-Seeded Ideation (Treatment). The teams were randomly assigned to two conditions: Control teams: five teams were provided with a report containing a set of risks brainstormed by AI practitioners in our first evaluation, i.e., blank-slate risk brainstorming workshops as described in Section 4.1; Treatment teams: six teams were instead provided with a report containing the same number of risks but generated by our framework with a small-world network topology (Section 3). All the risks were randomly sampled from the set of top prioritized risks from each ap- proach to ensure coverage of the wide spectrum. All teams then conducted the participatory AI risk assessment ideation exercise, brainstorming risks across three layers of conse- quences in the Futures Wheel. PER FUTURES WHEEL LAYER Layer 1 Layer 2 Layer 3 # of risks 0246810 PER MIT AI RISK TAXONOMY CATEGORY Discrimination and toxicity Privacy and security Misinformation Malicious actors Human-computer interaction Socioeconomic and environmental AI system safety failures and limitations 02468 AB # of risks AVERAGE NUMBER OF BRAINSTORMED RISKS IN ControlTreatment vs. (1 st order) (2 nd order) (3 rd order) Figure 4: Futures Wheel participatory ideation results. Teams seeded with our networked simulation generated risks (Treatment) identified more second- and third-order risks (A) than teams seeded with practitioner-generated risks (Control). They also surfaced a broader range of socio- technical harms, particularly human–computer interaction and socioeconomic and environmental risks (B). 5.2 Results: Our Framework Identifies More and Deeper Systemic Risks Overall, the treatment teams generated a substantially higher volume of risks compared to the control teams. On average, treatment teams brainstormed 21.1 risks per team, compared with 15.6 risks among control teams. Crucially, as shown in Figure 4A, this increase in volume was entirely driven by a greater capacity to uncover higher-order, systemic conse- quences, which tend to be harder to envision (Friedman and Hendry 2012). Both conditions yielded an identical average of 5.2 risks at Level 1 (immediate consequences). However, the treatment group significantly outperformed the control group in subsequent layers, identifying an average of 8.7 risks at Level 2 (compared to 5.2 in the control) and 7.2 risks at Level 3 (compared to 5.2 in the control). This demon- strates that our framework effectively prompts human partic- ipants to think deeper and anticipate cascading effects rather than exhausting their effort on obvious primary outcomes. To understand which risks were identified, we classified the brainstormed risks using the MIT AI Risk taxonomy (Slattery et al. 2026) (Figure 4B). We found that the treat- ment teams generated more risks across almost all cate- gories. Most notably, they surfaced substantially more risks in complex, downstream areas such as Human-Computer Interaction (8.2 vs. 6.0), Socioeconomic & Environmental harms (5.6 vs. 2.4), and AI System Safety Failures & Lim- itations (3.5 vs. 0.0). Interestingly, the control teams only produced more risks in the “Privacy & Security” category (1.5 vs. 1.0). This demonstrates practitioners-seeded base- lines often anchor human ideation on well-known, conven- tional risks, whereas the networked stakeholder simulation encourages the exploration of non-obvious, systemic harms. Ultimately, these results indicate that utilizing our frame- work to support the ideation stage of participatory AI risk assessments is significantly more effective than relying on a baseline of practitioner-generated risks. It not only increases the total number of identified risks, but specifically enhances human capacity to discover the higher-order socio-technical harms that are typically the hardest to spot. 6 Conclusion and Discussion We presented a framework that transforms AI risk assess- ment into a dynamic, multi-perspective exploration. By gen- erating stakeholder networks and utilizing social network analysis measures like Betweenness Centrality, we demon- strated a capability to uncover high-impact risks that tradi- tional methods miss. Our findings suggest that AI risk as- sessment benefits not only from who is involved, but from how stakeholder perspectives, even in-silico, are structured, connected, and weighted. By treating risk identification as a network-mediated ideation, we provide practitioners with a tool to help with a human-led ideation session in the partic- ipatory AI risk assessments. Implications. This work has three key implications: From flat independent assessments to network-mediated ideation. Current AI risk assessment practices, whether human-led brainstorming workshops or LLM-generated risk assessments, implicitly assume that risks are enumerable, independent, and equally visible to all evaluators. Risks emerge at the intersection of stakeholder roles (e.g., between legal compliance and user vulnerability, or between product incentives and mental health outcomes) are systematically underrepresented by flat generation methods. By modeling stakeholder interaction explicitly, our approach shifts the fo- cus from coverage of known risk categories to discovery of emergent, relational harms. Scalable pluralism and concrete usage pathways. Our framework is explicitly designed as a complementary pre- assessment tool, not a substitute for authentic participa- tion in participatory risk assessment. Risk management pro- fessionals can integrate this tool into early design sprints to conduct structured gap analyses against regulatory tax- onomies (e.g., the EU AI Act). By running the simulation prior to a human workshop, organizers can use the pre- generated materials, specifically areas of simulated disagree- ment and heavily populated systemic risks, as prompts for human ideation as we have done in our second evaluation. This prevents human participants from starting with a “blank slate” and focuses their valuable time contextualizing, and deepening complex risks. The tool allows organizers to ex- plicitly balance AI practitioners’ operational insights with the broader, theoretically informed views of researchers. If the simulation reveals that a specific bridging stakeholder (e.g., a “Mental Health Advocate”) is central to identify- ing high-severity risks, organizers are provided an evidence- based justification to invest resources into recruiting humans holding those lived experiences for subsequent participatory stages. We explicitly caution against using our tool to by- pass authentic inclusion with actual human participation for the full risk assessment. Most importantly, outputs from this tool must not be submitted to auditors or regulators as a sub- stitute for authentic stakeholder evidence or testimony. Prioritization as a normative choice. Our comparison of centrality network based ranking strategies highlights that risk prioritization is not value-neutral. Degree-based ranking favors widely recognized, consensus risks; Betweenness- based ranking elevates boundary-spanning concerns. This trade-off mirrors real-world governance tensions between addressing “known knowns” and anticipating low-frequency but high-impact harms (potentially to marginalized stake- holders). Making these prioritization logics explicit allows practitioners and regulators to align risk assessment outputs with their normative goals. Limitations. Our work has five main limitations. Firstly, our framework relies on LLM-simulated stake- holders to model structured information flow among network-conditioned agents, rather than authentic stake- holder deliberation (Habermas 1981). Because these in- silico representations cannot fully capture the lived expe- rience, contextual knowledge, or normative authority of real affected communities, certain harms, particularly those rooted in cultural, emotional, or marginalized perspectives, may be misrepresented or missed. Secondly, while our Futures Wheel study grounded the framework in a human-seeded task, it involved only one stakeholder type (young, non-Western users). Crucial do- main experts and vulnerable groups, such as mental health clinicians and social workers, were not included. System- atically benchmarking against workshops involving both highly vulnerable stakeholders and these specialized domain experts remains a critical direction for future work. Thirdly, the quality and diversity of generated risks de- pend on the underlying language model and the architec- tural choices made by the system designers. This introduces a structural paradox: while the tool is intended to mitigate the narrow biases of traditional human brainstorming, we (as researchers) defined the network topologies and prompt structures, which might inevitably encode our own biases into the system. Compounding this issue, because LLMs of- ten encode representational biases from historically skewed training data (Bai et al. 2025; Gupta et al. 2023; Shujaa, Dor- leon, and Tang 2025), simulated stakeholders run the risk of producing algorithmic caricatures rather than authentic reflections of marginalized communities (Vecchione, Levy, and Barocas 2021). Procedurally, it is imperative that our framework is strictly utilized as a pre-assessment tool to seed human ideation, not as a substitute for authentic human participation. A vital direction for future work is to system- atically benchmark these stereotyping effects and validate our simulations directly against workshops involving highly vulnerable, directly affected stakeholders to ensure the tool safely bridges, rather than widens, representational gaps. Fourthly, the constructed stakeholder networks encode simplifying assumptions about influence flow (e.g., acyclic- ity), which may not fully reflect the reciprocal and evolving nature of real-world stakeholder dynamics. Finally, our evaluation focuses on a single use case (AI chatbot companions), and further studies are needed to as- sess the generalizability of the approach across domains with different stakeholder structures and risk profiles. Ethics Statement All human participants, including the AI practitioners in the baseline study and the annotators recruited via Prolific, pro- vided informed consent prior to participation. No person- ally identifiable information was collected, and all data was anonymized prior to analysis. Participants in the baseline and Futures Wheel evaluation phases were compensated ex- ceeding the local minimum wage. For the expert verification tasks on Prolific, stringent screening criteria were applied to ensure domain expertise. References AI, N. 2023.Artificial Intelligence Risk Management framework (AI RMF 1.0). URL: https://nvlpubs. nist. gov- /nistpubs/ai/nist. ai, 100–1. AlDahoul, N.; Rahwan, T.; and Zaki, Y. 2025. Ai-Generated Faces Influence Gender Stereotypes and Racial Homoge- nization. Scientific Reports, 15(1). Ashkinaze, J.; Fry, E.; Edara, N.; Gilbert, E.; and Budak, C. 2025a. Plurals: A System for Guiding LLMs via Simulated Social Ensembles. In Proceedings of the ACM Conference on Human Factors in Computing Systems (CHI), 1–21. Ashkinaze, J.; Mendelsohn, J.; Qiwei, L.; Budak, C.; and Gilbert, E. 2025b. How AI Ideas Affect the Creativity, Di- versity, and Evolution of Human Ideas: Evidence From a Large, Dynamic Experiment. In Proceedings of the ACM Collective Intelligence Conference (CI), 198–213. Bai, X.; Wang, A.; Sucholutsky, I.; and Griffiths, T. L. 2025. Explicitly Unbiased Large Language Models Still Form Bi- ased Associations. Proceedings of the National Academy of Sciences, 122(8): e2416228122. Ballard, S.; Chappell, K. M.; and Kennedy, K. 2019. Judg- ment Call the Game: Using Value Sensitive Design and De- sign Fiction to Surface Ethical Concerns Related to Tech- nology. In Proceedings of the ACM Conference in Design- ing Interactive Systems Conference (DIS), 421–433. ACM. ISBN 9781450358507. Bengston, D. N. 2016. The Futures Wheel: A Method for Exploring the Implications of Social–Ecological Change. Society & natural resources, 29(3): 374–379. Birhane, A.; Isaac, W.; Prabhakaran, V.; Diaz, M.; Elish, M. C.; Gabriel, I.; and Mohamed, S. 2022. Power to the Peo- ple? Opportunities and Challenges for Participatory AI. In Proceedings of the ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO), 1–8. Borgatti, S. P.; Jones, C.; and Everett, M. G. 1998. Network Measures of Social Capital. Connections, 21(2): 27–36. Bouschery, S. G.; Blazevic, V.; and Piller, F. T. 2024. Arti- ficial Intelligence-Augmented Brainstorming: How Humans and AI Beat Humans Alone. SSRN Working Paper, posted April 3, 2024. Brooke, J.; et al. 1996. SUS: A Quick and Dirty Usability Scale. Usability Evaluation in Industry, 189(194): 4–7. Buc ̧inca, Z.; Pham, C. M.; Jakesch, M.; Ribeiro, M. T.; Olteanu, A.; and Amershi, S. 2023.AHA!: Facilitating AI Impact Assessment by Generating Examples of Harms. arXiv:2306.03280. Choi, D.; Hong, S.; Park, J.; Chung, J. J. Y.; and Kim, J. 2024. Creativeconnect: Supporting Reference Recombina- tion for Graphic Design Ideation With Generative AI. In Proceedings of the ACM Conference on Human Factors in Computing Systems (CHI), 1–25. Corbett, E.; Denton, R.; and Erete, S. 2023. Power and Pub- lic Participation in AI. In Proceedings of the ACM Confer- ence on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO), 1–13. Dean, D. L.; Hender, J. M.; Rodgers, T. L.; and Santanen, E. L. 2006. Identifying Quality, Novel, and Creative Ideas: Constructs and Scales for Idea Evaluation. Journal of the Association for Information Systems, 7(10): 646–699. Delgado, F.; Yang, S.; Madaio, M.; and Yang, Q. 2023. The Participatory Turn in AI Design: Theoretical Foundations and the Current State of Practice. In Proceedings of the ACM Conference on Equity and Access in Algorithms, Mecha- nisms, and Optimization (EAAMO), 1–23. Deng, W. H.; Claire, W.; Han, H. Z.; Hong, J. I.; Holstein, K.; and Eslami, M. 2025. WeAudit: Scaffolding User Audi- tors and AI Practitioners in Auditing Generative AI. Proc. ACM Hum.-Comput. Interact., 9(7). DeVos, A.; Dhabalia, A.; Shen, H.; Holstein, K.; and Es- lami, M. 2022. Toward User-Driven Algorithm Auditing: Investigating Users’ Strategies for Uncovering Harmful Al- gorithmic Behavior. In Proceedings of the ACM Conference on Human Factors in Computing Systems (CHI), 1–19. Diakopoulos, N.; and Johnson, D. 2021. Anticipating and Addressing the Ethical Implications of Deepfakes in the Context of Elections. New Media & Society, 23(7): 2072– 2098. Diehl, M.; and Stroebe, W. 1987. Productivity Loss in Brain- storming Groups: Toward the Solution of a Riddle. Journal of Personality and Social Psychology, 53(3): 497–509. European Commission. 2026. Consolidated text: Regula- tion (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Direc- tives 2014/90/EU, (EU) 2016/797 and (EU) 2020/1828 (Ar- tificial Intelligence Act) (Text with EEA relevance). Friedman, B.; and Hendry, D. 2012. The Envisioning Cards: A Toolkit for Catalyzing Humanistic and Technical Imagi- nations. In Proceedings of the SIGCHI Conference on Hu- man Factors in Computing Systems (CHI), 1145–1148. Fr ̈ ohling, L.; Giaconia, A.; Bogucka, E. P.; and Quercia, D. 2026. Agent-Supported Foresight for AI Systemic Risks: AI Agents for Breadth, Experts for Judgment. In Proceedings of the ACM Conference on Human Factors in Computing Systems (CHI). ACM. ISBN 9798400722783. Fukumura, K.; and Ito, T. 2025. Can LLM-Powered Multi- Agent Systems Augment Human Creativity? Evidence from Brainstorming Tasks.In Proceedings of the ACM Col- lective Intelligence Conference (CI), 20–29. ACM. ISBN 9798400714894. Ganguli, D.; Lovitt, L.; Kernion, J.; Askell, A.; Bai, Y.; Ka- davath, S.; Mann, B.; Perez, E.; Schiefer, N.; Ndousse, K.; et al. 2022.Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned. arXiv preprint arXiv:2209.07858. Ge, S.; Zhou, C.; Hou, R.; Khabsa, M.; Wang, Y.-C.; Wang, Q.; Han, J.; and Mao, Y. 2024. MART: Improving LLM Safety with Multi-round Automatic Red-Teaming. In Pro- ceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (ACL), 1927–1937. Glenn, J. C. 2009. The Futures Wheel. In Glenn, J. C.; and Gordon, T. J., eds., Futures Research Methodology, Version 3.0, 245–263. The Millennium Project. Govers, J.; ˇ S ́ cepanovi ́ c, S.; and Quercia, D. 2026. When and How AI Should Assist Brainstorming for AI Impact Assess- ment. In The 2026 ACM Conference on Fairness, Account- ability, and Transparency, 2745–2780. Gupta, S.; Shrivastava, V.; Deshpande, A.; Kalyan, A.; Clark, P.; Sabharwal, A.; and Khot, T. 2023.Bias Runs Deep: Implicit Reasoning Biases in Persona-Assigned LLms. arXiv preprint arXiv:2311.04892. Habermas, J. 1981. The Theory of Communicative Action. Beacon press. Herdel, V.; ˇ S ́ cepanovi ́ c, S.; Bogucka, E.; and Quercia, D. 2024. ExploreGen: Large Language Models for Envision- ing the Uses and Risks of AI Technologies. In Proceed- ings of the AAAI/ACM Conference on AI, Ethics, and Society (AIES), volume 7, 584–596. Hohendanner, M.; Ullstein, C.; Onyekwelu, B. A.; Katirai, A.; Kuribayashi, J.; Babalola, O.; Ema, A.; and Grossklags, J. 2025. Initiating the Global AI Dialogues: Laypeople Per- spectives on the Future Role of genAI in Society from Nige- ria, Germany and Japan. In Proceedings of the ACM Confer- ence on Human Factors in Computing Systems (CHI). ACM. ISBN 9798400713941. Hohma, E.; Boch, A.; Trauth, R.; and L ̈ utge, C. 2023. Inves- tigating Accountability for Artificial Intelligence Through Risk Governance: A Workshop-Based Exploratory Study. Frontiers in Psychology, 14: 1073686. Kallina, E.; Bohn ́ e, T.; and Singh, J. 2025. Stakeholder Participation for Responsible AI Development: Disconnects Between Guidance and Current Practice. In Proceedings of the ACM Conference on Fairness, Accountability, and Transparency (FAccT), 1060–1079. Kieslich, K.; Helberger, N.; and Diakopoulos, N. 2026. Scenario-based Sociotechnical Rnvisioning (SSE): An Ap- proach to Enhance Systemic Risk Assessments. AI and Ethics, 6(3). Lancaster, C. M.; Schulenberg, K.; Flathmann, C.; Mc- Neese, N. J.; and Freeman, G. 2024.“It’s Everybody’s Role to Speak Up... But Not Everyone Will”: Understand- ing AI Professionals’ Perceptions of Accountability for AI Bias Mitigation. ACM Journal on Responsible Computing, 1(1): 1–30. Lee, M. K.; Kusbit, D.; Kahng, A.; Kim, J. T.; Yuan, X.; Chan, A.; See, D.; Noothigattu, R.; Lee, S.; Psomas, A.; et al. 2019. WeBuildAI: Participatory Framework for Al- gorithmic Governance. Proceedings of the ACM on Human- Computer Interaction, 3(CSCW): 1–35. Lewis, J. R.; and Sauro, J. 2009. The Factor Structure of the System Usability Scale. In Kurosu, M., ed., Human Cen- tered Design, 94–103. Berlin, Heidelberg: Springer Berlin Heidelberg. ISBN 978-3-642-02806-9. Meincke, L.; Nave, G.; and Terwiesch, C. 2025. ChatGPT Decreases Idea Diversity in Brainstorming. Nature Human Behaviour, 9(6): 1107–1109. Park, J. S.; O’Brien, J.; Cai, C. J.; Morris, M. R.; Liang, P.; and Bernstein, M. S. 2023. Generative Agents: Interactive Simulacra of Human Behavior. In Proceedings of the An- nual ACM Symposium on User Interface Software and Tech- nology, 1–22. Parmar, B. L.; Freeman, R. E.; Harrison, J. S.; Wicks, A. C.; Purnell, L.; and De Colle, S. 2010. Stakeholder Theory: The State of the Art. Academy of Management Annals, 4(1): 403–445. Png, M.-T. 2022. At the Tensions of South and North: Crit- ical Roles of Global South Stakeholders in AI Governance. In Proceedings of the ACM Conference on Fairness, Ac- countability, and Transparency (FAccT), 1434–1445. Rowley, T. J. 1997. Moving Beyond Dyadic Ties: A Net- work Theory of Stakeholder Influences. Academy of Man- agement Review, 22(4): 887–910. San Vito, P. D. C.; Stumpf, S.; Hyde-Vaamonde, C.; and Thuermer, G. 2025. Ensuring Artificial Intelligence is Safe and Trustworthy: The Need for Participatory Auditing. In ACM Conference on Human Factors in Computing Systems (CHI): Sociotechnical AI Governance: Opportunities and Challenges for HCI. Scherr, S.; Cao, B.; Jiang, L. C.; and Kobayashi, T. 2025. Explaining the Use of AI Chatbots as Context Alignment: Motivations Behind the Use of AI Chatbots Across Contexts and Culture. Computers in Human Behavior, 172: 108738. Schmitz, A.; Mock, M.; G ̈ orge, R.; Cremers, A. B.; and Poretschkin, M. 2025. A Global Scale Comparison of Risk Aggregation in AI Assessment Frameworks. AI and Ethics, 5(2): 1407–1432. Serrano, M. ́ A.; Bogun ́ a, M.; and Vespignani, A. 2009. Ex- tracting the Multiscale Backbone of Complex Weighted Net- works. Proceedings of the National Academy of Sciences, 106(16): 6483–6488. Shujaa, S.; Dorleon, G.; and Tang, A. 2025. How LLMs Handle Cultural Bias: Reactions to Asian Minority Histori- cal Narratives. In International Conference on Asian Digital Libraries, 39–52. Springer. Slattery, P.; Saeri, A. K.; Grundy, E. A.; Graham, J.; Noetel, M.; Uuk, R.; Dao, J.; Pour, S.; Casper, S.; and Thompson, N. 2026. The AI Risk Repository: A Meta-Review, Database, and Taxonomy of Risks from Artificial Intelligence. Pat- terns, 7(5): 101517. Steimers, A.; and Schneider, M. 2022. Sources of Risk of AI Systems. International Journal of Environmental Research and Public Health, 19(6): 3641. Uuk, R.; Gutierrez, C. I.; Guppy, D.; Lauwaert, L.; Kasirzadeh, A.; Velasco, L.; Slattery, P.; and Prunkl, C. 2024.A Taxonomy of Systemic Risks From General- Purpose AI. arXiv:2412.07780. Vecchione, B.; Levy, K.; and Barocas, S. 2021. Algorith- mic Auditing and Social Justice: Lessons From the History of Audit Studies. In Proceedings of the ACM Conference on Equity and Access in Algorithms, Mechanisms, and Op- timization (EAAMO), 1–9. Wadinambiarachchi, S.; Kelly, R. M.; Pareek, S.; Zhou, Q.; and Velloso, E. 2024. The Effects of Generative AI on De- sign Fixation and Divergent Thinking. In Proceedings of the ACM Conference on Human Factors in Computing Systems (CHI). ACM. ISBN 9798400703300. Wang, W.-F.; Lu, C.-T.; Ponsa i Campany ` a, N.; Chen, B.-Y.; and Chen, M. Y. 2025. AIdeation: Designing a Human-AI Collaborative Ideation System for Concept Designers. In Proceedings of the ACM Conference on Human Factors in Computing Systems (CHI), 1–28. Wang, Z. J.; Kulkarni, C.; Wilcox, L.; Terry, M.; and Madaio, M. 2024.Farsight: Fostering Responsible AI Awareness During AI Application Prototyping. In Proceed- ings of the ACM Conference on Human Factors in Comput- ing Systems (CHI), 1–40. Watts, D. J.; and Strogatz, S. H. 1998. Collective Dynamics of ‘Small-World’ Networks. Nature, 393(6684): 440–442. Weidinger, L.; Uesato, J.; Rauh, M.; Griffin, C.; Huang, P.- S.; Mellor, J.; Glaese, A.; Cheng, M.; Balle, B.; Kasirzadeh, A.; et al. 2022. Taxonomy of Risks Posed by Language Models. In Proceedings of the ACM Conference on Fair- ness, Accountability, and Transparency (FAccT), 214–229. Wilson, C. 2018. Is It Love or Loneliness? Exploring the Im- pact of Everyday Digital Technology Use on the Wellbeing of Older Adults. Ageing & Society, 38(7): 1307–1331. Xia, B.; Lu, Q.; Perera, H.; Zhu, L.; Xing, Z.; Liu, Y.; and Whittle, J. 2023. Towards Concrete and Connected AI Risk Assessment (C2AIRA): A Systematic Mapping Study. In IEEE/ACM International Conference on AI Engineering– Software Engineering for AI (CAIN), 104–116. IEEE Com- puter Society. Yu-Han, C.; and Chun-Ching, C. 2023. Investigating the Im- pact of Generative Artificial Intelligence on Brainstorming: A Preliminary Study. In International Conference on Con- sumer Electronics - Taiwan (ICCE-Taiwan), 193–194. Zhang, R.; Li, H.; Meng, H.; Zhan, J.; Gan, H.; and Lee, Y.- C. 2025. The Dark Side of AI Companionship: A Taxonomy of Harmful Algorithmic Behaviors in Human-AI Relation- ships. In Proceedings of the ACM Conference on Human Factors in Computing Systems (CHI), 1–17. A Appendix A.1 Implementations To operationalize the complex interactions required for our stakeholder simulation, we adopt and extend the Plurals framework proposed by Ashkinaze et al. (Ashkinaze et al. 2025a). Plurals is a modular system designed to steer Large Language Models via simulated recursive social deliberation. It abstracts the chaotic process of multi-agent conversation into three governable primitives, which we map to our risk assessment task: Agents (The Stakeholders): In Plurals, agents are LLM instances initialized with specific “system prompts” and “personas”. We map each node s in our stakeholder graph to a Plurals Agent, injecting the specific role description (e.g., “software engi- neer”) into their initialization state. Structures (The Topology): The framework manages information flow through defined topologies (e.g., Ensembles, Debates, Chains). While the original framework focuses on static interaction patterns, we extend this by dynamically feeding our con- structed network structures (Hierarchical, Small-World, Random) into the Plurals graph executor. This ensures that an Agent only receives context from its specific neighbors defined in our network generation phase (i.e., local connected network), rather than a global context window. Moderators (The Synthesizers): These are specialized agents tasked with processing the output of a deliberation round. We utilize Moderators not to force consensus, as is common in political deliberation tasks, but to perform deduplication and formatting of risks between rounds, preventing the “echo chamber” effect where agents merely repeat their neighbors’ inputs. Our implementation utilizes the Plurals Library (https://josh-ashkinaze.github.io/plurals/) to handle the asynchronous or- chestration of these agents, while our contribution lies in the dynamic generation of the Structure (via snowballing) and the post-hoc analysis of the graph using network measures for prioritizing risks. To address the known representational biases in LLMs, we implemented targeted prompt diversification. Agent prompts explicitly instructed the LLMs to adopt an exploratory and uncertainty-aware framing (e.g., “Identify potential socio-technical risks that *might* uniquely affect this stakeholder group”). Moderators were prompted to synthesize without erasing minority dissenting views, counteracting the LLM tendency to collapse into consensus stereotypes. A.2 Supplementary Experimental Results Table 2 summarizes the performance of all system variants across network structures, stakeholder strategies, and ranking mea- sures based on two metrics: (1)“Combined Score” that averages Plausibility, Likelihood, Uniqueness, Engagement and Severity, and (2) “Novelty” that solely reflects risk diversity. All measures were normalized to five-point Likert scale. Generic Seeding Enables Broader Stakeholder Discovery We examine the initialization strategy: is it better to manually curate a nuanced list of stakeholders, or to seed the system with a few generic roles and allow it to “snowball” (recursively discover) new ones? We focus this analysis using Betweenness Centrality (Table 3), as this measure best captures the value of discovering bridging stakeholders. We find that starting with a Typical (generic) stakeholder set and utilizing snowballing is remarkably effective. As shown in Table 3, the “Typical + Snowballing” configuration achieves a higher Combined Score (3.99) than the “Nuanced + Snowballing” configuration (3.94), while maintaining a relatively high Novelty score (3.48 vs 3.42). This is a counter-intuitive but significant finding. It suggests that providing a nuanced list ex ante may induce path dependency, constraining the LLM to a pre-defined search space and limiting the “organic” discovery of bridging nodes. In contrast, generic seeding acts as a “minimum viable prompt”. When the system is allowed to snowball from a simple seed, it identifies not just more stakeholders, but better connected ones—agents that bridge gaps between the core roles. Consequently, practitioners do not need to invest heavily in curating exhaustive stakeholder lists. A simple seed list, combined with a dynamic network expansion process and betweenness-based prioritization, is sufficient to uncover high-quality, novel risks that outperform static nuanced lists. A.3 Prolific Expert and Rubric Validation To ensure high-quality annotations for our comparison, we recruited five external AI-domain experts via Prolific. We applied strict screening criteria: participants were required to have a professional background in AI ethics assessment, and pass both an initial comprehension check and embedded attention checks during the task. They were compensated above minimum wage. To validate our rubric dimensions, we computed the Pearson correlation among these scores across all evaluated risks (Figure A.3). The low-to-moderate correlations among most of these metrics confirm that these dimensions capture distinct evaluative constructs rather than redundant signals (although as expected, Likelihood is more strongly correlated with Plausibility). Nov- elty is marginally negatively correlated with Likelihood and Uniqueness, while weakly positively correlated with all other measures. Table 2: Performance of AI risk assessment system variants, organized by network variant (structure + measure). “Combined Score” aggregates Plausibility, Likelihood, Uniqueness, Engagement and Severity while “Novelty” solely reflects risk diversity. All measures were normalized to the five-point Likert scale, and the top-two performing systems are bolded. We observe that performance improves systematically with (i) small-world interaction structures, (i) dynamic stakeholder expansion, and (i) network-aware prioritization. Degree centrality yields the highest combined risk scores, whereas betweenness centrality surfaces the most novel and diverse risks. Network Variant (Structure + Measure)Roles (Stakeholder Strategy)Combined ScoreNovelty Baselines NoneSingle-turn LLM3.292.32 Implicit DAGMulti-turn Plurals3.722.98 Hierarchical Network HierarchyFixed Typical Roles (5 roles)3.152.21 HierarchyFixed Comprehensive Roles3.582.70 HierarchySnowballing Typical Roles3.732.95 HierarchySnowballing Comprehensive Roles3.773.00 Small-World Network + Degree Centrality Small-world + DegreeFixed Typical Roles (5 roles)3.262.17 Small-world + DegreeFixed Comprehensive Roles3.782.95 Small-world + DegreeSnowballing Typical Roles4.053.18 Small-world + DegreeSnowballing Comprehensive Roles4.013.22 Small-World Network + Betweenness Centrality Small-world + BetweennessFixed Typical Roles (5 roles)3.192.36 Small-world + BetweennessFixed Comprehensive Roles3.733.12 Small-world + BetweennessSnowballing Typical Roles3.993.48 Small-world + BetweennessSnowballing Comprehensive Roles3.943.42 Small-World Network + Closeness Centrality Small-world + ClosenessFixed Typical Roles (5 roles)3.122.32 Small-world + ClosenessFixed Comprehensive Roles3.672.78 Small-world + ClosenessSnowballing Typical Roles3.963.02 Small-world + ClosenessSnowballing Comprehensive Roles3.913.06 Random Network + Degree Centrality Random + DegreeFixed Typical Roles (5 roles)3.142.33 Random + DegreeFixed Comprehensive Roles3.582.86 Random + DegreeSnowballing Typical Roles3.903.05 Random + DegreeSnowballing Comprehensive Roles3.823.10 Random Network + Betweenness Centrality Random + BetweennessFixed Typical Roles (5 roles)3.072.38 Random + BetweennessFixed Comprehensive Roles3.513.02 Random + BetweennessSnowballing Typical Roles3.783.22 Random + BetweennessSnowballing Comprehensive Roles3.753.28 Random Network + Closeness Centrality Random + ClosenessFixed Typical Roles (5 roles)3.062.19 Random + ClosenessFixed Comprehensive Roles3.482.72 Random + ClosenessSnowballing Typical Roles3.762.95 Random + ClosenessSnowballing Comprehensive Roles3.683.00 Table 3: Effect of Seeding Strategy on Risk Prioritization (Small-World + Betweenness Centrality). Dynamic expansion (Snow- balling) drastically improves performance over fixed stakeholders. Notably, snowballing from a small, generic set (Typical) slightly outperforms starting with a more specific set (Nuanced) on Combined Score (3.99 vs 3.94) and Novelty (3.48 vs 3.42), indicating that organic network growth effectively discovers critical bridging roles. Initial SeedExpansion StrategyCombinedNovelty Typical (5 generic stakeholders)Fixed (No expansion)3.192.36 Nuanced (30 specific stakeholders)Fixed (No expansion)3.733.12 Typical (5 generic stakeholders)Snowballing3.993.48 Nuanced (30 specific stakeholders)Snowballing3.943.42 Table 4: Demographics of participants in the Chatbot Companion workshops for our first evaluation (blank slate human gener- ated risks). CharacteristicNumber of Participants (%) Gender Men15 (33.3%) Women30 (66.7%) Age 18–248 (17.8%) 25–3423 (51.1%) 35–448 (17.8%) 45+1 (2.2%) Education No degree0 (0.0%) Bachelor’s degree8 (17.8%) Master’s degree23 (51.1%) PhD9 (20.0%) AI Expertise Basic5 (11.1%) General7 (15.6%) Knowledgeable17 (37.8%) Expert12 (26.7%) Field of Work/Study CS / AI / Data Science27 (60.0%) Health & Life Sciences6 (13.3%) Physical Sciences & Engineering2 (4.4%) Humanities & Ethics3 (6.7%) Business3 (6.7%) Others4 (8.9%) Total participants45 A.4 Risk Impact Assessment Reports Figure 5: Risk impact assessment report provided to participant groups in the control condition of the second evaluation. The report contains risks generated by AI practitioners in blank-slate brainstorming workshops (Section 4.1), organized into technical, human interaction, and systemic risk categories. Figure 6: Risk impact assessment report provided to participant groups in the treatment condition of the second eval- uation. The report contains risks generated by our networked stakeholder simulation framework (Section 3), organized into technical, human interaction, and systemic risk categories. Figure 7: Correlation matrix of various rubric dimensions demonstrates that these dimensions capture distinct evaluative signals rather than redundant measures. Novelty is marginally negatively correlated with Likelihood and Uniqueness, while weakly positively correlated with all other measures.

Instant research discovery